10 #include <effect_chain.h>
11 #include <effect_util.h>
12 #include <epoxy/egl.h>
14 #include <image_format.h>
16 #include <overlay_effect.h>
17 #include <padding_effect.h>
18 #include <resample_effect.h>
19 #include <resource_pool.h>
20 #include <saturation_effect.h>
27 #include <white_balance_effect.h>
29 #include <ycbcr_input.h>
31 #include <condition_variable>
41 #include "h264encode.h"
42 #include "pbo_frame_allocator.h"
43 #include "ref_counted_gl_sync.h"
47 using namespace movit;
49 using namespace std::placeholders;
51 Mixer *global_mixer = nullptr;
53 Mixer::Mixer(const QSurfaceFormat &format)
54 : mixer_surface(create_surface(format)),
55 h264_encoder_surface(create_surface(format))
57 CHECK(init_movit(MOVIT_SHADER_DIR, MOVIT_DEBUG_OFF));
60 resource_pool.reset(new ResourcePool);
61 theme.reset(new Theme("theme.lua", resource_pool.get()));
62 output_channel[OUTPUT_LIVE].parent = this;
63 output_channel[OUTPUT_PREVIEW].parent = this;
64 output_channel[OUTPUT_INPUT0].parent = this;
65 output_channel[OUTPUT_INPUT1].parent = this;
67 ImageFormat inout_format;
68 inout_format.color_space = COLORSPACE_sRGB;
69 inout_format.gamma_curve = GAMMA_sRGB;
71 YCbCrFormat input_ycbcr_format;
72 input_ycbcr_format.chroma_subsampling_x = 2;
73 input_ycbcr_format.chroma_subsampling_y = 1;
74 input_ycbcr_format.cb_x_position = 0.0;
75 input_ycbcr_format.cr_x_position = 0.0;
76 input_ycbcr_format.cb_y_position = 0.5;
77 input_ycbcr_format.cr_y_position = 0.5;
78 input_ycbcr_format.luma_coefficients = YCBCR_REC_601;
79 input_ycbcr_format.full_range = false;
81 // Display chain; shows the live output produced by the main chain (its RGBA version).
82 display_chain.reset(new EffectChain(WIDTH, HEIGHT, resource_pool.get()));
84 display_input = new FlatInput(inout_format, FORMAT_RGB, GL_UNSIGNED_BYTE, WIDTH, HEIGHT); // FIXME: GL_UNSIGNED_BYTE is really wrong.
85 display_chain->add_input(display_input);
86 display_chain->add_output(inout_format, OUTPUT_ALPHA_FORMAT_POSTMULTIPLIED);
87 display_chain->set_dither_bits(0); // Don't bother.
88 display_chain->finalize();
90 // Preview chains (always shows just the inputs for now).
91 preview0_chain.reset(new EffectChain(WIDTH, HEIGHT, resource_pool.get()));
93 preview0_input = new YCbCrInput(inout_format, input_ycbcr_format, WIDTH, HEIGHT, YCBCR_INPUT_SPLIT_Y_AND_CBCR);
94 preview0_chain->add_input(preview0_input);
95 preview0_chain->add_output(inout_format, OUTPUT_ALPHA_FORMAT_POSTMULTIPLIED);
96 preview0_chain->set_dither_bits(0); // Don't bother.
97 preview0_chain->finalize();
99 preview1_chain.reset(new EffectChain(WIDTH, HEIGHT, resource_pool.get()));
101 preview1_input = new YCbCrInput(inout_format, input_ycbcr_format, WIDTH, HEIGHT, YCBCR_INPUT_SPLIT_Y_AND_CBCR);
102 preview1_chain->add_input(preview1_input);
103 preview1_chain->add_output(inout_format, OUTPUT_ALPHA_FORMAT_POSTMULTIPLIED);
104 preview1_chain->set_dither_bits(0); // Don't bother.
105 preview1_chain->finalize();
107 h264_encoder.reset(new H264Encoder(h264_encoder_surface, WIDTH, HEIGHT, "test.mp4"));
109 printf("Configuring first card...\n");
110 cards[0].usb = new BMUSBCapture(0x1edb, 0xbd3b); // 0xbd4f
111 cards[0].usb->set_frame_callback(std::bind(&Mixer::bm_frame, this, 0, _1, _2, _3, _4, _5, _6, _7));
112 cards[0].frame_allocator.reset(new PBOFrameAllocator(1280 * 750 * 2 + 44, 1280, 720));
113 cards[0].usb->set_video_frame_allocator(cards[0].frame_allocator.get());
114 cards[0].usb->configure_card();
115 cards[0].surface = create_surface(format);
117 cards[1].surface = create_surface(format);
120 if (NUM_CARDS == 2) {
121 printf("Configuring second card...\n");
122 cards[1].usb = new BMUSBCapture(0x1edb, 0xbd4f);
123 cards[1].usb->set_frame_callback(std::bind(&Mixer::bm_frame, this, 1, _1, _2, _3, _4, _5, _6, _7));
124 cards[1].frame_allocator.reset(new PBOFrameAllocator(1280 * 750 * 2 + 44, 1280, 720));
125 cards[1].usb->set_video_frame_allocator(cards[1].frame_allocator.get());
126 cards[1].usb->configure_card();
129 BMUSBCapture::start_bm_thread();
131 for (int card_index = 0; card_index < NUM_CARDS; ++card_index) {
132 cards[card_index].usb->start_bm_capture();
135 //chain->enable_phase_timing(true);
137 // Set up stuff for NV12 conversion.
140 string cbcr_vert_shader = read_file("vs-cbcr.130.vert");
141 string cbcr_frag_shader =
144 "uniform sampler2D cbcr_tex; \n"
146 " gl_FragColor = texture2D(cbcr_tex, tc0); \n"
148 cbcr_program_num = resource_pool->compile_glsl_program(cbcr_vert_shader, cbcr_frag_shader);
153 resource_pool->release_glsl_program(cbcr_program_num);
154 BMUSBCapture::stop_bm_thread();
157 void Mixer::bm_frame(int card_index, uint16_t timecode,
158 FrameAllocator::Frame video_frame, size_t video_offset, uint16_t video_format,
159 FrameAllocator::Frame audio_frame, size_t audio_offset, uint16_t audio_format)
161 CaptureCard *card = &cards[card_index];
162 if (!card->thread_initialized) {
163 printf("initializing context for bmusb thread %d\n", card_index);
164 eglBindAPI(EGL_OPENGL_API);
165 card->context = create_context();
166 if (!make_current(card->context, card->surface)) {
167 printf("failed to create bmusb context\n");
170 card->thread_initialized = true;
173 if (video_frame.len - video_offset != 1280 * 750 * 2) {
174 printf("dropping frame with wrong length (%ld)\n", video_frame.len - video_offset);
175 FILE *fp = fopen("frame.raw", "wb");
176 fwrite(video_frame.data, video_frame.len, 1, fp);
179 card->usb->get_video_frame_allocator()->release_frame(video_frame);
180 card->usb->get_audio_frame_allocator()->release_frame(audio_frame);
184 // Wait until the previous frame was consumed.
185 std::unique_lock<std::mutex> lock(bmusb_mutex);
186 card->new_data_ready_changed.wait(lock, [card]{ return !card->new_data_ready; });
188 const PBOFrameAllocator::Userdata *userdata = (const PBOFrameAllocator::Userdata *)video_frame.userdata;
189 GLuint pbo = userdata->pbo;
191 glBindBuffer(GL_PIXEL_UNPACK_BUFFER_ARB, pbo);
193 glFlushMappedBufferRange(GL_PIXEL_UNPACK_BUFFER, 0, video_frame.size);
195 //glMemoryBarrier(GL_CLIENT_MAPPED_BUFFER_BARRIER_BIT);
198 // Upload the textures.
199 glBindTexture(GL_TEXTURE_2D, userdata->tex_y);
201 glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, 1280, 720, GL_RED, GL_UNSIGNED_BYTE, BUFFER_OFFSET((1280 * 750 * 2 + 44) / 2 + 1280 * 25 + 22));
203 glBindTexture(GL_TEXTURE_2D, userdata->tex_cbcr);
205 glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, 1280/2, 720, GL_RG, GL_UNSIGNED_BYTE, BUFFER_OFFSET(1280 * 25 + 22));
207 glBindTexture(GL_TEXTURE_2D, 0);
209 GLsync fence = glFenceSync(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
211 assert(fence != nullptr);
213 std::unique_lock<std::mutex> lock(bmusb_mutex);
214 card->new_data_ready = true;
215 card->new_frame = RefCountedFrame(video_frame);
216 card->new_data_ready_fence = fence;
217 card->new_data_ready_changed.notify_all();
220 // Video frame will be released when last user of card->new_frame goes out of scope.
221 card->usb->get_audio_frame_allocator()->release_frame(audio_frame);
224 void Mixer::place_rectangle(Effect *resample_effect, Effect *padding_effect, float x0, float y0, float x1, float y1)
232 if (x0 > 1280.0 || x1 < 0.0 || y0 > 720.0 || y1 < 0.0) {
233 CHECK(resample_effect->set_int("width", 1));
234 CHECK(resample_effect->set_int("height", 1));
235 CHECK(resample_effect->set_float("zoom_x", 1280.0));
236 CHECK(resample_effect->set_float("zoom_y", 720.0));
237 CHECK(padding_effect->set_int("left", 2000));
238 CHECK(padding_effect->set_int("top", 2000));
242 // Clip. (TODO: Clip on upper/left sides, too.)
244 srcx1 = (1280.0 - x0) / (x1 - x0);
248 srcy1 = (720.0 - y0) / (y1 - y0);
252 float x_subpixel_offset = x0 - floor(x0);
253 float y_subpixel_offset = y0 - floor(y0);
255 // Resampling must be to an integral number of pixels. Round up,
256 // and then add an extra pixel so we have some leeway for the border.
257 int width = int(ceil(x1 - x0)) + 1;
258 int height = int(ceil(y1 - y0)) + 1;
259 CHECK(resample_effect->set_int("width", width));
260 CHECK(resample_effect->set_int("height", height));
262 // Correct the discrepancy with zoom. (This will leave a small
263 // excess edge of pixels and subpixels, which we'll correct for soon.)
264 float zoom_x = (x1 - x0) / (width * (srcx1 - srcx0));
265 float zoom_y = (y1 - y0) / (height * (srcy1 - srcy0));
266 CHECK(resample_effect->set_float("zoom_x", zoom_x));
267 CHECK(resample_effect->set_float("zoom_y", zoom_y));
268 CHECK(resample_effect->set_float("zoom_center_x", 0.0f));
269 CHECK(resample_effect->set_float("zoom_center_y", 0.0f));
271 // Padding must also be to a whole-pixel offset.
272 CHECK(padding_effect->set_int("left", floor(x0)));
273 CHECK(padding_effect->set_int("top", floor(y0)));
275 // Correct _that_ discrepancy by subpixel offset in the resampling.
276 CHECK(resample_effect->set_float("left", -x_subpixel_offset / zoom_x));
277 CHECK(resample_effect->set_float("top", -y_subpixel_offset / zoom_y));
279 // Finally, adjust the border so it is exactly where we want it.
280 CHECK(padding_effect->set_float("border_offset_left", x_subpixel_offset));
281 CHECK(padding_effect->set_float("border_offset_right", x1 - (floor(x0) + width)));
282 CHECK(padding_effect->set_float("border_offset_top", y_subpixel_offset));
283 CHECK(padding_effect->set_float("border_offset_bottom", y1 - (floor(y0) + height)));
286 void Mixer::thread_func()
288 eglBindAPI(EGL_OPENGL_API);
289 QOpenGLContext *context = create_context();
290 if (!make_current(context, mixer_surface)) {
295 struct timespec start, now;
296 clock_gettime(CLOCK_MONOTONIC, &start);
298 while (!should_quit) {
301 //int width0 = lrintf(848 * (1.0 + 0.2 * sin(frame * 0.02)));
303 int height0 = lrintf(width0 * 9.0 / 16.0);
305 //float top0 = 96 + 48 * sin(frame * 0.005);
306 //float left0 = 96 + 48 * cos(frame * 0.006);
309 float bottom0 = top0 + height0;
310 float right0 = left0 + width0;
315 float bottom1 = 720 - 48;
316 float right1 = 1280 - 16;
317 float top1 = bottom1 - height1;
318 float left1 = right1 - width1;
320 if (current_source == SOURCE_INPUT1) {
327 bottom1 = HEIGHT + 20;
330 } else if (current_source == SOURCE_INPUT2) {
337 bottom0 = HEIGHT + 20;
341 float t = 0.5 + 0.5 * cos(frame * 0.006);
342 float scale0 = 1.0 + t * (1280.0 / 848.0 - 1.0);
343 float tx0 = 0.0 + t * (-16.0 * scale0);
344 float ty0 = 0.0 + t * (-48.0 * scale0);
346 top0 = top0 * scale0 + ty0;
347 bottom0 = bottom0 * scale0 + ty0;
348 left0 = left0 * scale0 + tx0;
349 right0 = right0 * scale0 + tx0;
351 top1 = top1 * scale0 + ty0;
352 bottom1 = bottom1 * scale0 + ty0;
353 left1 = left1 * scale0 + tx0;
354 right1 = right1 * scale0 + tx0;
358 place_rectangle(resample_effect, padding_effect, left0, top0, right0, bottom0);
359 place_rectangle(resample2_effect, padding2_effect, left1, top1, right1, bottom1);
362 CaptureCard card_copy[NUM_CARDS];
365 std::unique_lock<std::mutex> lock(bmusb_mutex);
367 // The first card is the master timer, so wait for it to have a new frame.
368 // TODO: Make configurable, and with a timeout.
369 cards[0].new_data_ready_changed.wait(lock, [this]{ return cards[0].new_data_ready; });
371 for (int card_index = 0; card_index < NUM_CARDS; ++card_index) {
372 CaptureCard *card = &cards[card_index];
373 card_copy[card_index].usb = card->usb;
374 card_copy[card_index].new_data_ready = card->new_data_ready;
375 card_copy[card_index].new_frame = card->new_frame;
376 card_copy[card_index].new_data_ready_fence = card->new_data_ready_fence;
377 card->new_data_ready = false;
378 card->new_data_ready_changed.notify_all();
382 for (int card_index = 0; card_index < NUM_CARDS; ++card_index) {
383 CaptureCard *card = &card_copy[card_index];
384 if (!card->new_data_ready)
387 assert(card->new_frame != nullptr);
388 bmusb_current_rendering_frame[card_index] = card->new_frame;
391 // The new texture might still be uploaded,
392 // tell the GPU to wait until it's there.
393 if (card->new_data_ready_fence)
394 glWaitSync(card->new_data_ready_fence, /*flags=*/0, GL_TIMEOUT_IGNORED);
396 glDeleteSync(card->new_data_ready_fence);
398 const PBOFrameAllocator::Userdata *userdata = (const PBOFrameAllocator::Userdata *)card->new_frame->userdata;
399 theme->set_input_textures(card_index, userdata->tex_y, userdata->tex_cbcr);
402 // Get the main chain from the theme, and set its state immediately.
403 pair<EffectChain *, function<void()>> theme_main_chain = theme->get_chain(0, frame / 60.0f, WIDTH, HEIGHT);
404 EffectChain *chain = theme_main_chain.first;
405 theme_main_chain.second();
407 GLuint y_tex, cbcr_tex;
408 bool got_frame = h264_encoder->begin_frame(&y_tex, &cbcr_tex);
411 // Render main chain.
412 GLuint cbcr_full_tex = resource_pool->create_2d_texture(GL_RG8, WIDTH, HEIGHT);
413 GLuint rgba_tex = resource_pool->create_2d_texture(GL_RGB565, WIDTH, HEIGHT); // Saves texture bandwidth, although dithering gets messed up.
414 GLuint fbo = resource_pool->create_fbo(y_tex, cbcr_full_tex, rgba_tex);
415 chain->render_to_fbo(fbo, WIDTH, HEIGHT);
416 resource_pool->release_fbo(fbo);
418 subsample_chroma(cbcr_full_tex, cbcr_tex);
419 resource_pool->release_2d_texture(cbcr_full_tex);
421 // Set the right state for rgba_tex.
422 glBindFramebuffer(GL_FRAMEBUFFER, 0);
423 glBindTexture(GL_TEXTURE_2D, rgba_tex);
424 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
425 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
426 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
428 RefCountedGLsync fence(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
431 // Make sure the H.264 gets a reference to all the
432 // input frames needed, so that they are not released back
433 // until the rendering is done.
434 vector<RefCountedFrame> input_frames;
435 for (int card_index = 0; card_index < NUM_CARDS; ++card_index) {
436 input_frames.push_back(bmusb_current_rendering_frame[card_index]);
438 h264_encoder->end_frame(fence, input_frames);
440 // The live frame just shows the RGBA texture we just rendered.
441 // It owns rgba_tex now.
442 DisplayFrame live_frame;
443 live_frame.chain = display_chain.get();
444 live_frame.setup_chain = [this, rgba_tex]{
445 display_input->set_texture_num(rgba_tex);
447 live_frame.ready_fence = fence;
448 live_frame.input_frames = {};
449 live_frame.temp_textures = { rgba_tex };
450 output_channel[OUTPUT_LIVE].output_frame(live_frame);
452 // The preview frame shows the first input. Note that the textures
453 // are owned by the input frame, not the display frame.
454 if (bmusb_current_rendering_frame[0] != nullptr) {
455 const PBOFrameAllocator::Userdata *input0_userdata = (const PBOFrameAllocator::Userdata *)bmusb_current_rendering_frame[0]->userdata;
456 GLuint input0_y_tex = input0_userdata->tex_y;
457 GLuint input0_cbcr_tex = input0_userdata->tex_cbcr;
458 DisplayFrame preview0_frame;
459 preview0_frame.chain = preview0_chain.get();
460 preview0_frame.setup_chain = [this, input0_y_tex, input0_cbcr_tex]{
461 preview0_input->set_texture_num(0, input0_y_tex);
462 preview0_input->set_texture_num(1, input0_cbcr_tex);
464 preview0_frame.ready_fence = fence;
465 preview0_frame.input_frames = { bmusb_current_rendering_frame[0] };
466 preview0_frame.temp_textures = {};
467 output_channel[OUTPUT_PREVIEW].output_frame(preview0_frame);
468 output_channel[OUTPUT_INPUT0].output_frame(preview0_frame);
471 // Same for the other preview.
472 // TODO: Use a for loop. Gah.
473 if (bmusb_current_rendering_frame[1] != nullptr) {
474 const PBOFrameAllocator::Userdata *input1_userdata = (const PBOFrameAllocator::Userdata *)bmusb_current_rendering_frame[1]->userdata;
475 GLuint input1_y_tex = input1_userdata->tex_y;
476 GLuint input1_cbcr_tex = input1_userdata->tex_cbcr;
477 DisplayFrame preview1_frame;
478 preview1_frame.chain = preview1_chain.get();
479 preview1_frame.setup_chain = [this, input1_y_tex, input1_cbcr_tex]{
480 preview1_input->set_texture_num(0, input1_y_tex);
481 preview1_input->set_texture_num(1, input1_cbcr_tex);
483 preview1_frame.ready_fence = fence;
484 preview1_frame.input_frames = { bmusb_current_rendering_frame[1] };
485 preview1_frame.temp_textures = {};
486 output_channel[OUTPUT_INPUT1].output_frame(preview1_frame);
489 clock_gettime(CLOCK_MONOTONIC, &now);
490 double elapsed = now.tv_sec - start.tv_sec +
491 1e-9 * (now.tv_nsec - start.tv_nsec);
492 if (frame % 100 == 0) {
493 printf("%d frames in %.3f seconds = %.1f fps (%.1f ms/frame)\n",
494 frame, elapsed, frame / elapsed,
495 1e3 * elapsed / frame);
496 // chain->print_phase_timing();
499 // Reset every 100 frames, so that local variations in frame times
500 // (especially for the first few frames, when the shaders are
501 // compiled etc.) don't make it hard to measure for the entire
502 // remaining duration of the program.
503 if (frame == 10000) {
511 void Mixer::subsample_chroma(GLuint src_tex, GLuint dst_tex)
514 glGenVertexArrays(1, &vao);
523 glBindVertexArray(vao);
527 GLuint fbo = resource_pool->create_fbo(dst_tex);
528 glBindFramebuffer(GL_FRAMEBUFFER, fbo);
529 glViewport(0, 0, WIDTH/2, HEIGHT/2);
532 glUseProgram(cbcr_program_num);
535 glActiveTexture(GL_TEXTURE0);
537 glBindTexture(GL_TEXTURE_2D, src_tex);
539 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
541 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
543 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
546 float chroma_offset_0[] = { -0.5f / WIDTH, 0.0f };
547 set_uniform_vec2(cbcr_program_num, "foo", "chroma_offset_0", chroma_offset_0);
549 GLuint position_vbo = fill_vertex_attribute(cbcr_program_num, "position", 2, GL_FLOAT, sizeof(vertices), vertices);
550 GLuint texcoord_vbo = fill_vertex_attribute(cbcr_program_num, "texcoord", 2, GL_FLOAT, sizeof(vertices), vertices); // Same as vertices.
552 glDrawArrays(GL_TRIANGLES, 0, 3);
555 cleanup_vertex_attribute(cbcr_program_num, "position", position_vbo);
556 cleanup_vertex_attribute(cbcr_program_num, "texcoord", texcoord_vbo);
561 resource_pool->release_fbo(fbo);
562 glDeleteVertexArrays(1, &vao);
565 void Mixer::release_display_frame(DisplayFrame *frame)
567 for (GLuint texnum : frame->temp_textures) {
568 resource_pool->release_2d_texture(texnum);
570 frame->temp_textures.clear();
571 frame->ready_fence.reset();
572 frame->input_frames.clear();
577 mixer_thread = std::thread(&Mixer::thread_func, this);
586 void Mixer::cut(Source source)
588 current_source = source;
591 void Mixer::OutputChannel::output_frame(DisplayFrame frame)
593 // Store this frame for display. Remove the ready frame if any
594 // (it was seemingly never used).
596 std::unique_lock<std::mutex> lock(frame_mutex);
597 if (has_ready_frame) {
598 parent->release_display_frame(&ready_frame);
601 has_ready_frame = true;
604 if (has_new_frame_ready_callback) {
605 new_frame_ready_callback();
609 bool Mixer::OutputChannel::get_display_frame(DisplayFrame *frame)
611 std::unique_lock<std::mutex> lock(frame_mutex);
612 if (!has_current_frame && !has_ready_frame) {
616 if (has_current_frame && has_ready_frame) {
617 // We have a new ready frame. Toss the current one.
618 parent->release_display_frame(¤t_frame);
619 has_current_frame = false;
621 if (has_ready_frame) {
622 assert(!has_current_frame);
623 current_frame = ready_frame;
624 ready_frame.ready_fence.reset(); // Drop the refcount.
625 ready_frame.input_frames.clear(); // Drop the refcounts.
626 has_current_frame = true;
627 has_ready_frame = false;
630 *frame = current_frame;
634 void Mixer::OutputChannel::set_frame_ready_callback(Mixer::new_frame_ready_callback_t callback)
636 new_frame_ready_callback = callback;
637 has_new_frame_ready_callback = true;